2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
20 * - Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials
23 * provided with the distribution.
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 * $Id: ipoib_main.c 1377 2004-12-23 19:57:12Z roland $
39 #include <linux/module.h>
41 #include <linux/init.h>
42 #include <linux/slab.h>
43 #include <linux/kernel.h>
44 #include <linux/vmalloc.h>
46 #include <linux/if_arp.h> /* For ARPHRD_xxx */
53 MODULE_AUTHOR("Roland Dreier");
54 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
55 MODULE_LICENSE("Dual BSD/GPL");
57 int ipoib_sendq_size __read_mostly
= IPOIB_TX_RING_SIZE
;
58 int ipoib_recvq_size __read_mostly
= IPOIB_RX_RING_SIZE
;
60 module_param_named(send_queue_size
, ipoib_sendq_size
, int, 0444);
61 MODULE_PARM_DESC(send_queue_size
, "Number of descriptors in send queue");
62 module_param_named(recv_queue_size
, ipoib_recvq_size
, int, 0444);
63 MODULE_PARM_DESC(recv_queue_size
, "Number of descriptors in receive queue");
65 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
66 int ipoib_debug_level
;
68 module_param_named(debug_level
, ipoib_debug_level
, int, 0644);
69 MODULE_PARM_DESC(debug_level
, "Enable debug tracing if > 0");
72 struct ipoib_path_iter
{
73 struct net_device
*dev
;
74 struct ipoib_path path
;
77 static const u8 ipv4_bcast_addr
[] = {
78 0x00, 0xff, 0xff, 0xff,
79 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
80 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
83 struct workqueue_struct
*ipoib_workqueue
;
85 struct ib_sa_client ipoib_sa_client
;
87 static void ipoib_add_one(struct ib_device
*device
);
88 static void ipoib_remove_one(struct ib_device
*device
);
90 static struct ib_client ipoib_client
= {
93 .remove
= ipoib_remove_one
96 int ipoib_open(struct net_device
*dev
)
98 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
100 ipoib_dbg(priv
, "bringing up interface\n");
102 napi_enable(&priv
->napi
);
103 set_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
);
105 if (ipoib_pkey_dev_delay_open(dev
))
108 if (ipoib_ib_dev_open(dev
)) {
109 napi_disable(&priv
->napi
);
113 if (ipoib_ib_dev_up(dev
)) {
114 ipoib_ib_dev_stop(dev
, 1);
115 napi_disable(&priv
->napi
);
119 if (!test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
)) {
120 struct ipoib_dev_priv
*cpriv
;
122 /* Bring up any child interfaces too */
123 mutex_lock(&priv
->vlan_mutex
);
124 list_for_each_entry(cpriv
, &priv
->child_intfs
, list
) {
127 flags
= cpriv
->dev
->flags
;
131 dev_change_flags(cpriv
->dev
, flags
| IFF_UP
);
133 mutex_unlock(&priv
->vlan_mutex
);
136 netif_start_queue(dev
);
141 static int ipoib_stop(struct net_device
*dev
)
143 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
145 ipoib_dbg(priv
, "stopping interface\n");
147 clear_bit(IPOIB_FLAG_ADMIN_UP
, &priv
->flags
);
148 napi_disable(&priv
->napi
);
150 netif_stop_queue(dev
);
153 * Now flush workqueue to make sure a scheduled task doesn't
154 * bring our internal state back up.
156 flush_workqueue(ipoib_workqueue
);
158 ipoib_ib_dev_down(dev
, 1);
159 ipoib_ib_dev_stop(dev
, 1);
161 if (!test_bit(IPOIB_FLAG_SUBINTERFACE
, &priv
->flags
)) {
162 struct ipoib_dev_priv
*cpriv
;
164 /* Bring down any child interfaces too */
165 mutex_lock(&priv
->vlan_mutex
);
166 list_for_each_entry(cpriv
, &priv
->child_intfs
, list
) {
169 flags
= cpriv
->dev
->flags
;
170 if (!(flags
& IFF_UP
))
173 dev_change_flags(cpriv
->dev
, flags
& ~IFF_UP
);
175 mutex_unlock(&priv
->vlan_mutex
);
181 static int ipoib_change_mtu(struct net_device
*dev
, int new_mtu
)
183 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
185 /* dev->mtu > 2K ==> connected mode */
186 if (ipoib_cm_admin_enabled(dev
)) {
187 if (new_mtu
> ipoib_cm_max_mtu(dev
))
190 if (new_mtu
> priv
->mcast_mtu
)
191 ipoib_warn(priv
, "mtu > %d will cause multicast packet drops.\n",
198 if (new_mtu
> IPOIB_UD_MTU(priv
->max_ib_mtu
))
201 priv
->admin_mtu
= new_mtu
;
203 dev
->mtu
= min(priv
->mcast_mtu
, priv
->admin_mtu
);
208 static struct ipoib_path
*__path_find(struct net_device
*dev
, void *gid
)
210 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
211 struct rb_node
*n
= priv
->path_tree
.rb_node
;
212 struct ipoib_path
*path
;
216 path
= rb_entry(n
, struct ipoib_path
, rb_node
);
218 ret
= memcmp(gid
, path
->pathrec
.dgid
.raw
,
219 sizeof (union ib_gid
));
232 static int __path_add(struct net_device
*dev
, struct ipoib_path
*path
)
234 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
235 struct rb_node
**n
= &priv
->path_tree
.rb_node
;
236 struct rb_node
*pn
= NULL
;
237 struct ipoib_path
*tpath
;
242 tpath
= rb_entry(pn
, struct ipoib_path
, rb_node
);
244 ret
= memcmp(path
->pathrec
.dgid
.raw
, tpath
->pathrec
.dgid
.raw
,
245 sizeof (union ib_gid
));
254 rb_link_node(&path
->rb_node
, pn
, n
);
255 rb_insert_color(&path
->rb_node
, &priv
->path_tree
);
257 list_add_tail(&path
->list
, &priv
->path_list
);
262 static void path_free(struct net_device
*dev
, struct ipoib_path
*path
)
264 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
265 struct ipoib_neigh
*neigh
, *tn
;
269 while ((skb
= __skb_dequeue(&path
->queue
)))
270 dev_kfree_skb_irq(skb
);
272 spin_lock_irqsave(&priv
->lock
, flags
);
274 list_for_each_entry_safe(neigh
, tn
, &path
->neigh_list
, list
) {
276 * It's safe to call ipoib_put_ah() inside priv->lock
277 * here, because we know that path->ah will always
278 * hold one more reference, so ipoib_put_ah() will
279 * never do more than decrement the ref count.
282 ipoib_put_ah(neigh
->ah
);
284 ipoib_neigh_free(dev
, neigh
);
287 spin_unlock_irqrestore(&priv
->lock
, flags
);
290 ipoib_put_ah(path
->ah
);
295 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
297 struct ipoib_path_iter
*ipoib_path_iter_init(struct net_device
*dev
)
299 struct ipoib_path_iter
*iter
;
301 iter
= kmalloc(sizeof *iter
, GFP_KERNEL
);
306 memset(iter
->path
.pathrec
.dgid
.raw
, 0, 16);
308 if (ipoib_path_iter_next(iter
)) {
316 int ipoib_path_iter_next(struct ipoib_path_iter
*iter
)
318 struct ipoib_dev_priv
*priv
= netdev_priv(iter
->dev
);
320 struct ipoib_path
*path
;
323 spin_lock_irq(&priv
->lock
);
325 n
= rb_first(&priv
->path_tree
);
328 path
= rb_entry(n
, struct ipoib_path
, rb_node
);
330 if (memcmp(iter
->path
.pathrec
.dgid
.raw
, path
->pathrec
.dgid
.raw
,
331 sizeof (union ib_gid
)) < 0) {
340 spin_unlock_irq(&priv
->lock
);
345 void ipoib_path_iter_read(struct ipoib_path_iter
*iter
,
346 struct ipoib_path
*path
)
351 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
353 void ipoib_flush_paths(struct net_device
*dev
)
355 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
356 struct ipoib_path
*path
, *tp
;
357 LIST_HEAD(remove_list
);
359 spin_lock_irq(&priv
->tx_lock
);
360 spin_lock(&priv
->lock
);
362 list_splice_init(&priv
->path_list
, &remove_list
);
364 list_for_each_entry(path
, &remove_list
, list
)
365 rb_erase(&path
->rb_node
, &priv
->path_tree
);
367 list_for_each_entry_safe(path
, tp
, &remove_list
, list
) {
369 ib_sa_cancel_query(path
->query_id
, path
->query
);
370 spin_unlock(&priv
->lock
);
371 spin_unlock_irq(&priv
->tx_lock
);
372 wait_for_completion(&path
->done
);
373 path_free(dev
, path
);
374 spin_lock_irq(&priv
->tx_lock
);
375 spin_lock(&priv
->lock
);
377 spin_unlock(&priv
->lock
);
378 spin_unlock_irq(&priv
->tx_lock
);
381 static void path_rec_completion(int status
,
382 struct ib_sa_path_rec
*pathrec
,
385 struct ipoib_path
*path
= path_ptr
;
386 struct net_device
*dev
= path
->dev
;
387 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
388 struct ipoib_ah
*ah
= NULL
;
389 struct ipoib_neigh
*neigh
, *tn
;
390 struct sk_buff_head skqueue
;
395 ipoib_dbg(priv
, "PathRec LID 0x%04x for GID " IPOIB_GID_FMT
"\n",
396 be16_to_cpu(pathrec
->dlid
), IPOIB_GID_ARG(pathrec
->dgid
));
398 ipoib_dbg(priv
, "PathRec status %d for GID " IPOIB_GID_FMT
"\n",
399 status
, IPOIB_GID_ARG(path
->pathrec
.dgid
));
401 skb_queue_head_init(&skqueue
);
404 struct ib_ah_attr av
;
406 if (!ib_init_ah_from_path(priv
->ca
, priv
->port
, pathrec
, &av
))
407 ah
= ipoib_create_ah(dev
, priv
->pd
, &av
);
410 spin_lock_irqsave(&priv
->lock
, flags
);
415 path
->pathrec
= *pathrec
;
417 ipoib_dbg(priv
, "created address handle %p for LID 0x%04x, SL %d\n",
418 ah
, be16_to_cpu(pathrec
->dlid
), pathrec
->sl
);
420 while ((skb
= __skb_dequeue(&path
->queue
)))
421 __skb_queue_tail(&skqueue
, skb
);
423 list_for_each_entry_safe(neigh
, tn
, &path
->neigh_list
, list
) {
424 kref_get(&path
->ah
->ref
);
425 neigh
->ah
= path
->ah
;
426 memcpy(&neigh
->dgid
.raw
, &path
->pathrec
.dgid
.raw
,
427 sizeof(union ib_gid
));
429 if (ipoib_cm_enabled(dev
, neigh
->neighbour
)) {
430 if (!ipoib_cm_get(neigh
))
431 ipoib_cm_set(neigh
, ipoib_cm_create_tx(dev
,
434 if (!ipoib_cm_get(neigh
)) {
435 list_del(&neigh
->list
);
437 ipoib_put_ah(neigh
->ah
);
438 ipoib_neigh_free(dev
, neigh
);
443 while ((skb
= __skb_dequeue(&neigh
->queue
)))
444 __skb_queue_tail(&skqueue
, skb
);
449 complete(&path
->done
);
451 spin_unlock_irqrestore(&priv
->lock
, flags
);
453 while ((skb
= __skb_dequeue(&skqueue
))) {
455 if (dev_queue_xmit(skb
))
456 ipoib_warn(priv
, "dev_queue_xmit failed "
457 "to requeue packet\n");
461 static struct ipoib_path
*path_rec_create(struct net_device
*dev
, void *gid
)
463 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
464 struct ipoib_path
*path
;
466 if (!priv
->broadcast
)
469 path
= kzalloc(sizeof *path
, GFP_ATOMIC
);
475 skb_queue_head_init(&path
->queue
);
477 INIT_LIST_HEAD(&path
->neigh_list
);
479 memcpy(path
->pathrec
.dgid
.raw
, gid
, sizeof (union ib_gid
));
480 path
->pathrec
.sgid
= priv
->local_gid
;
481 path
->pathrec
.pkey
= cpu_to_be16(priv
->pkey
);
482 path
->pathrec
.numb_path
= 1;
483 path
->pathrec
.traffic_class
= priv
->broadcast
->mcmember
.traffic_class
;
488 static int path_rec_start(struct net_device
*dev
,
489 struct ipoib_path
*path
)
491 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
493 ipoib_dbg(priv
, "Start path record lookup for " IPOIB_GID_FMT
"\n",
494 IPOIB_GID_ARG(path
->pathrec
.dgid
));
496 init_completion(&path
->done
);
499 ib_sa_path_rec_get(&ipoib_sa_client
, priv
->ca
, priv
->port
,
501 IB_SA_PATH_REC_DGID
|
502 IB_SA_PATH_REC_SGID
|
503 IB_SA_PATH_REC_NUMB_PATH
|
504 IB_SA_PATH_REC_TRAFFIC_CLASS
|
509 if (path
->query_id
< 0) {
510 ipoib_warn(priv
, "ib_sa_path_rec_get failed\n");
512 return path
->query_id
;
518 static void neigh_add_path(struct sk_buff
*skb
, struct net_device
*dev
)
520 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
521 struct ipoib_path
*path
;
522 struct ipoib_neigh
*neigh
;
524 neigh
= ipoib_neigh_alloc(skb
->dst
->neighbour
, skb
->dev
);
526 ++dev
->stats
.tx_dropped
;
527 dev_kfree_skb_any(skb
);
532 * We can only be called from ipoib_start_xmit, so we're
533 * inside tx_lock -- no need to save/restore flags.
535 spin_lock(&priv
->lock
);
537 path
= __path_find(dev
, skb
->dst
->neighbour
->ha
+ 4);
539 path
= path_rec_create(dev
, skb
->dst
->neighbour
->ha
+ 4);
543 __path_add(dev
, path
);
546 list_add_tail(&neigh
->list
, &path
->neigh_list
);
549 kref_get(&path
->ah
->ref
);
550 neigh
->ah
= path
->ah
;
551 memcpy(&neigh
->dgid
.raw
, &path
->pathrec
.dgid
.raw
,
552 sizeof(union ib_gid
));
554 if (ipoib_cm_enabled(dev
, neigh
->neighbour
)) {
555 if (!ipoib_cm_get(neigh
))
556 ipoib_cm_set(neigh
, ipoib_cm_create_tx(dev
, path
, neigh
));
557 if (!ipoib_cm_get(neigh
)) {
558 list_del(&neigh
->list
);
560 ipoib_put_ah(neigh
->ah
);
561 ipoib_neigh_free(dev
, neigh
);
564 if (skb_queue_len(&neigh
->queue
) < IPOIB_MAX_PATH_REC_QUEUE
)
565 __skb_queue_tail(&neigh
->queue
, skb
);
567 ipoib_warn(priv
, "queue length limit %d. Packet drop.\n",
568 skb_queue_len(&neigh
->queue
));
572 ipoib_send(dev
, skb
, path
->ah
, IPOIB_QPN(skb
->dst
->neighbour
->ha
));
576 if (!path
->query
&& path_rec_start(dev
, path
))
579 __skb_queue_tail(&neigh
->queue
, skb
);
582 spin_unlock(&priv
->lock
);
586 list_del(&neigh
->list
);
589 ipoib_neigh_free(dev
, neigh
);
591 ++dev
->stats
.tx_dropped
;
592 dev_kfree_skb_any(skb
);
594 spin_unlock(&priv
->lock
);
597 static void ipoib_path_lookup(struct sk_buff
*skb
, struct net_device
*dev
)
599 struct ipoib_dev_priv
*priv
= netdev_priv(skb
->dev
);
601 /* Look up path record for unicasts */
602 if (skb
->dst
->neighbour
->ha
[4] != 0xff) {
603 neigh_add_path(skb
, dev
);
607 /* Add in the P_Key for multicasts */
608 skb
->dst
->neighbour
->ha
[8] = (priv
->pkey
>> 8) & 0xff;
609 skb
->dst
->neighbour
->ha
[9] = priv
->pkey
& 0xff;
610 ipoib_mcast_send(dev
, skb
->dst
->neighbour
->ha
+ 4, skb
);
613 static void unicast_arp_send(struct sk_buff
*skb
, struct net_device
*dev
,
614 struct ipoib_pseudoheader
*phdr
)
616 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
617 struct ipoib_path
*path
;
620 * We can only be called from ipoib_start_xmit, so we're
621 * inside tx_lock -- no need to save/restore flags.
623 spin_lock(&priv
->lock
);
625 path
= __path_find(dev
, phdr
->hwaddr
+ 4);
627 path
= path_rec_create(dev
, phdr
->hwaddr
+ 4);
629 /* put pseudoheader back on for next time */
630 skb_push(skb
, sizeof *phdr
);
631 __skb_queue_tail(&path
->queue
, skb
);
633 if (path_rec_start(dev
, path
)) {
634 spin_unlock(&priv
->lock
);
635 path_free(dev
, path
);
638 __path_add(dev
, path
);
640 ++dev
->stats
.tx_dropped
;
641 dev_kfree_skb_any(skb
);
644 spin_unlock(&priv
->lock
);
649 ipoib_dbg(priv
, "Send unicast ARP to %04x\n",
650 be16_to_cpu(path
->pathrec
.dlid
));
652 ipoib_send(dev
, skb
, path
->ah
, IPOIB_QPN(phdr
->hwaddr
));
653 } else if ((path
->query
|| !path_rec_start(dev
, path
)) &&
654 skb_queue_len(&path
->queue
) < IPOIB_MAX_PATH_REC_QUEUE
) {
655 /* put pseudoheader back on for next time */
656 skb_push(skb
, sizeof *phdr
);
657 __skb_queue_tail(&path
->queue
, skb
);
659 ++dev
->stats
.tx_dropped
;
660 dev_kfree_skb_any(skb
);
663 spin_unlock(&priv
->lock
);
666 static int ipoib_start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
668 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
669 struct ipoib_neigh
*neigh
;
672 if (unlikely(!spin_trylock_irqsave(&priv
->tx_lock
, flags
)))
673 return NETDEV_TX_LOCKED
;
675 if (likely(skb
->dst
&& skb
->dst
->neighbour
)) {
676 if (unlikely(!*to_ipoib_neigh(skb
->dst
->neighbour
))) {
677 ipoib_path_lookup(skb
, dev
);
681 neigh
= *to_ipoib_neigh(skb
->dst
->neighbour
);
684 if (unlikely((memcmp(&neigh
->dgid
.raw
,
685 skb
->dst
->neighbour
->ha
+ 4,
686 sizeof(union ib_gid
))) ||
687 (neigh
->dev
!= dev
))) {
688 spin_lock(&priv
->lock
);
690 * It's safe to call ipoib_put_ah() inside
691 * priv->lock here, because we know that
692 * path->ah will always hold one more reference,
693 * so ipoib_put_ah() will never do more than
694 * decrement the ref count.
696 ipoib_put_ah(neigh
->ah
);
697 list_del(&neigh
->list
);
698 ipoib_neigh_free(dev
, neigh
);
699 spin_unlock(&priv
->lock
);
700 ipoib_path_lookup(skb
, dev
);
704 if (ipoib_cm_get(neigh
)) {
705 if (ipoib_cm_up(neigh
)) {
706 ipoib_cm_send(dev
, skb
, ipoib_cm_get(neigh
));
709 } else if (neigh
->ah
) {
710 ipoib_send(dev
, skb
, neigh
->ah
, IPOIB_QPN(skb
->dst
->neighbour
->ha
));
714 if (skb_queue_len(&neigh
->queue
) < IPOIB_MAX_PATH_REC_QUEUE
) {
715 spin_lock(&priv
->lock
);
716 __skb_queue_tail(&neigh
->queue
, skb
);
717 spin_unlock(&priv
->lock
);
719 ++dev
->stats
.tx_dropped
;
720 dev_kfree_skb_any(skb
);
723 struct ipoib_pseudoheader
*phdr
=
724 (struct ipoib_pseudoheader
*) skb
->data
;
725 skb_pull(skb
, sizeof *phdr
);
727 if (phdr
->hwaddr
[4] == 0xff) {
728 /* Add in the P_Key for multicast*/
729 phdr
->hwaddr
[8] = (priv
->pkey
>> 8) & 0xff;
730 phdr
->hwaddr
[9] = priv
->pkey
& 0xff;
732 ipoib_mcast_send(dev
, phdr
->hwaddr
+ 4, skb
);
734 /* unicast GID -- should be ARP or RARP reply */
736 if ((be16_to_cpup((__be16
*) skb
->data
) != ETH_P_ARP
) &&
737 (be16_to_cpup((__be16
*) skb
->data
) != ETH_P_RARP
)) {
738 ipoib_warn(priv
, "Unicast, no %s: type %04x, QPN %06x "
740 skb
->dst
? "neigh" : "dst",
741 be16_to_cpup((__be16
*) skb
->data
),
742 IPOIB_QPN(phdr
->hwaddr
),
743 IPOIB_GID_RAW_ARG(phdr
->hwaddr
+ 4));
744 dev_kfree_skb_any(skb
);
745 ++dev
->stats
.tx_dropped
;
749 unicast_arp_send(skb
, dev
, phdr
);
754 spin_unlock_irqrestore(&priv
->tx_lock
, flags
);
759 static void ipoib_timeout(struct net_device
*dev
)
761 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
763 ipoib_warn(priv
, "transmit timeout: latency %d msecs\n",
764 jiffies_to_msecs(jiffies
- dev
->trans_start
));
765 ipoib_warn(priv
, "queue stopped %d, tx_head %u, tx_tail %u\n",
766 netif_queue_stopped(dev
),
767 priv
->tx_head
, priv
->tx_tail
);
768 /* XXX reset QP, etc. */
771 static int ipoib_hard_header(struct sk_buff
*skb
,
772 struct net_device
*dev
,
774 const void *daddr
, const void *saddr
, unsigned len
)
776 struct ipoib_header
*header
;
778 header
= (struct ipoib_header
*) skb_push(skb
, sizeof *header
);
780 header
->proto
= htons(type
);
781 header
->reserved
= 0;
784 * If we don't have a neighbour structure, stuff the
785 * destination address onto the front of the skb so we can
786 * figure out where to send the packet later.
788 if ((!skb
->dst
|| !skb
->dst
->neighbour
) && daddr
) {
789 struct ipoib_pseudoheader
*phdr
=
790 (struct ipoib_pseudoheader
*) skb_push(skb
, sizeof *phdr
);
791 memcpy(phdr
->hwaddr
, daddr
, INFINIBAND_ALEN
);
797 static void ipoib_set_mcast_list(struct net_device
*dev
)
799 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
801 if (!test_bit(IPOIB_FLAG_OPER_UP
, &priv
->flags
)) {
802 ipoib_dbg(priv
, "IPOIB_FLAG_OPER_UP not set");
806 queue_work(ipoib_workqueue
, &priv
->restart_task
);
809 static void ipoib_neigh_cleanup(struct neighbour
*n
)
811 struct ipoib_neigh
*neigh
;
812 struct ipoib_dev_priv
*priv
= netdev_priv(n
->dev
);
814 struct ipoib_ah
*ah
= NULL
;
816 neigh
= *to_ipoib_neigh(n
);
818 priv
= netdev_priv(neigh
->dev
);
822 "neigh_cleanup for %06x " IPOIB_GID_FMT
"\n",
824 IPOIB_GID_RAW_ARG(n
->ha
+ 4));
826 spin_lock_irqsave(&priv
->lock
, flags
);
830 list_del(&neigh
->list
);
831 ipoib_neigh_free(n
->dev
, neigh
);
833 spin_unlock_irqrestore(&priv
->lock
, flags
);
839 struct ipoib_neigh
*ipoib_neigh_alloc(struct neighbour
*neighbour
,
840 struct net_device
*dev
)
842 struct ipoib_neigh
*neigh
;
844 neigh
= kmalloc(sizeof *neigh
, GFP_ATOMIC
);
848 neigh
->neighbour
= neighbour
;
850 *to_ipoib_neigh(neighbour
) = neigh
;
851 skb_queue_head_init(&neigh
->queue
);
852 ipoib_cm_set(neigh
, NULL
);
857 void ipoib_neigh_free(struct net_device
*dev
, struct ipoib_neigh
*neigh
)
860 *to_ipoib_neigh(neigh
->neighbour
) = NULL
;
861 while ((skb
= __skb_dequeue(&neigh
->queue
))) {
862 ++dev
->stats
.tx_dropped
;
863 dev_kfree_skb_any(skb
);
865 if (ipoib_cm_get(neigh
))
866 ipoib_cm_destroy_tx(ipoib_cm_get(neigh
));
870 static int ipoib_neigh_setup_dev(struct net_device
*dev
, struct neigh_parms
*parms
)
872 parms
->neigh_cleanup
= ipoib_neigh_cleanup
;
877 int ipoib_dev_init(struct net_device
*dev
, struct ib_device
*ca
, int port
)
879 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
881 /* Allocate RX/TX "rings" to hold queued skbs */
882 priv
->rx_ring
= kzalloc(ipoib_recvq_size
* sizeof *priv
->rx_ring
,
884 if (!priv
->rx_ring
) {
885 printk(KERN_WARNING
"%s: failed to allocate RX ring (%d entries)\n",
886 ca
->name
, ipoib_recvq_size
);
890 priv
->tx_ring
= vmalloc(ipoib_sendq_size
* sizeof *priv
->tx_ring
);
891 if (!priv
->tx_ring
) {
892 printk(KERN_WARNING
"%s: failed to allocate TX ring (%d entries)\n",
893 ca
->name
, ipoib_sendq_size
);
894 goto out_rx_ring_cleanup
;
896 memset(priv
->tx_ring
, 0, ipoib_sendq_size
* sizeof *priv
->tx_ring
);
898 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
900 if (ipoib_ib_dev_init(dev
, ca
, port
))
901 goto out_tx_ring_cleanup
;
906 vfree(priv
->tx_ring
);
909 kfree(priv
->rx_ring
);
915 void ipoib_dev_cleanup(struct net_device
*dev
)
917 struct ipoib_dev_priv
*priv
= netdev_priv(dev
), *cpriv
, *tcpriv
;
919 ipoib_delete_debug_files(dev
);
921 /* Delete any child interfaces first */
922 list_for_each_entry_safe(cpriv
, tcpriv
, &priv
->child_intfs
, list
) {
923 unregister_netdev(cpriv
->dev
);
924 ipoib_dev_cleanup(cpriv
->dev
);
925 free_netdev(cpriv
->dev
);
928 ipoib_ib_dev_cleanup(dev
);
930 kfree(priv
->rx_ring
);
931 vfree(priv
->tx_ring
);
933 priv
->rx_ring
= NULL
;
934 priv
->tx_ring
= NULL
;
937 static const struct header_ops ipoib_header_ops
= {
938 .create
= ipoib_hard_header
,
941 static void ipoib_setup(struct net_device
*dev
)
943 struct ipoib_dev_priv
*priv
= netdev_priv(dev
);
945 dev
->open
= ipoib_open
;
946 dev
->stop
= ipoib_stop
;
947 dev
->change_mtu
= ipoib_change_mtu
;
948 dev
->hard_start_xmit
= ipoib_start_xmit
;
949 dev
->tx_timeout
= ipoib_timeout
;
950 dev
->header_ops
= &ipoib_header_ops
;
951 dev
->set_multicast_list
= ipoib_set_mcast_list
;
952 dev
->neigh_setup
= ipoib_neigh_setup_dev
;
954 ipoib_set_ethtool_ops(dev
);
956 netif_napi_add(dev
, &priv
->napi
, ipoib_poll
, 100);
958 dev
->watchdog_timeo
= HZ
;
960 dev
->flags
|= IFF_BROADCAST
| IFF_MULTICAST
;
963 * We add in INFINIBAND_ALEN to allow for the destination
964 * address "pseudoheader" for skbs without neighbour struct.
966 dev
->hard_header_len
= IPOIB_ENCAP_LEN
+ INFINIBAND_ALEN
;
967 dev
->addr_len
= INFINIBAND_ALEN
;
968 dev
->type
= ARPHRD_INFINIBAND
;
969 dev
->tx_queue_len
= ipoib_sendq_size
* 2;
970 dev
->features
= (NETIF_F_VLAN_CHALLENGED
|
974 memcpy(dev
->broadcast
, ipv4_bcast_addr
, INFINIBAND_ALEN
);
976 netif_carrier_off(dev
);
980 spin_lock_init(&priv
->lock
);
981 spin_lock_init(&priv
->tx_lock
);
983 mutex_init(&priv
->mcast_mutex
);
984 mutex_init(&priv
->vlan_mutex
);
986 INIT_LIST_HEAD(&priv
->path_list
);
987 INIT_LIST_HEAD(&priv
->child_intfs
);
988 INIT_LIST_HEAD(&priv
->dead_ahs
);
989 INIT_LIST_HEAD(&priv
->multicast_list
);
991 INIT_DELAYED_WORK(&priv
->pkey_poll_task
, ipoib_pkey_poll
);
992 INIT_WORK(&priv
->pkey_event_task
, ipoib_pkey_event
);
993 INIT_DELAYED_WORK(&priv
->mcast_task
, ipoib_mcast_join_task
);
994 INIT_WORK(&priv
->flush_task
, ipoib_ib_dev_flush
);
995 INIT_WORK(&priv
->restart_task
, ipoib_mcast_restart_task
);
996 INIT_DELAYED_WORK(&priv
->ah_reap_task
, ipoib_reap_ah
);
999 struct ipoib_dev_priv
*ipoib_intf_alloc(const char *name
)
1001 struct net_device
*dev
;
1003 dev
= alloc_netdev((int) sizeof (struct ipoib_dev_priv
), name
,
1008 return netdev_priv(dev
);
1011 static ssize_t
show_pkey(struct device
*dev
,
1012 struct device_attribute
*attr
, char *buf
)
1014 struct ipoib_dev_priv
*priv
= netdev_priv(to_net_dev(dev
));
1016 return sprintf(buf
, "0x%04x\n", priv
->pkey
);
1018 static DEVICE_ATTR(pkey
, S_IRUGO
, show_pkey
, NULL
);
1020 static ssize_t
show_umcast(struct device
*dev
,
1021 struct device_attribute
*attr
, char *buf
)
1023 struct ipoib_dev_priv
*priv
= netdev_priv(to_net_dev(dev
));
1025 return sprintf(buf
, "%d\n", test_bit(IPOIB_FLAG_UMCAST
, &priv
->flags
));
1028 static ssize_t
set_umcast(struct device
*dev
,
1029 struct device_attribute
*attr
,
1030 const char *buf
, size_t count
)
1032 struct ipoib_dev_priv
*priv
= netdev_priv(to_net_dev(dev
));
1033 unsigned long umcast_val
= simple_strtoul(buf
, NULL
, 0);
1035 if (umcast_val
> 0) {
1036 set_bit(IPOIB_FLAG_UMCAST
, &priv
->flags
);
1037 ipoib_warn(priv
, "ignoring multicast groups joined directly "
1040 clear_bit(IPOIB_FLAG_UMCAST
, &priv
->flags
);
1044 static DEVICE_ATTR(umcast
, S_IWUSR
| S_IRUGO
, show_umcast
, set_umcast
);
1046 int ipoib_add_umcast_attr(struct net_device
*dev
)
1048 return device_create_file(&dev
->dev
, &dev_attr_umcast
);
1051 static ssize_t
create_child(struct device
*dev
,
1052 struct device_attribute
*attr
,
1053 const char *buf
, size_t count
)
1058 if (sscanf(buf
, "%i", &pkey
) != 1)
1061 if (pkey
< 0 || pkey
> 0xffff)
1065 * Set the full membership bit, so that we join the right
1066 * broadcast group, etc.
1070 ret
= ipoib_vlan_add(to_net_dev(dev
), pkey
);
1072 return ret
? ret
: count
;
1074 static DEVICE_ATTR(create_child
, S_IWUGO
, NULL
, create_child
);
1076 static ssize_t
delete_child(struct device
*dev
,
1077 struct device_attribute
*attr
,
1078 const char *buf
, size_t count
)
1083 if (sscanf(buf
, "%i", &pkey
) != 1)
1086 if (pkey
< 0 || pkey
> 0xffff)
1089 ret
= ipoib_vlan_delete(to_net_dev(dev
), pkey
);
1091 return ret
? ret
: count
;
1094 static DEVICE_ATTR(delete_child
, S_IWUGO
, NULL
, delete_child
);
1096 int ipoib_add_pkey_attr(struct net_device
*dev
)
1098 return device_create_file(&dev
->dev
, &dev_attr_pkey
);
1101 static struct net_device
*ipoib_add_port(const char *format
,
1102 struct ib_device
*hca
, u8 port
)
1104 struct ipoib_dev_priv
*priv
;
1105 struct ib_device_attr
*device_attr
;
1106 struct ib_port_attr attr
;
1107 int result
= -ENOMEM
;
1109 priv
= ipoib_intf_alloc(format
);
1111 goto alloc_mem_failed
;
1113 SET_NETDEV_DEV(priv
->dev
, hca
->dma_device
);
1115 if (!ib_query_port(hca
, port
, &attr
))
1116 priv
->max_ib_mtu
= ib_mtu_enum_to_int(attr
.max_mtu
);
1118 printk(KERN_WARNING
"%s: ib_query_port %d failed\n",
1120 goto device_init_failed
;
1123 /* MTU will be reset when mcast join happens */
1124 priv
->dev
->mtu
= IPOIB_UD_MTU(priv
->max_ib_mtu
);
1125 priv
->mcast_mtu
= priv
->admin_mtu
= priv
->dev
->mtu
;
1127 result
= ib_query_pkey(hca
, port
, 0, &priv
->pkey
);
1129 printk(KERN_WARNING
"%s: ib_query_pkey port %d failed (ret = %d)\n",
1130 hca
->name
, port
, result
);
1131 goto device_init_failed
;
1134 device_attr
= kmalloc(sizeof *device_attr
, GFP_KERNEL
);
1136 printk(KERN_WARNING
"%s: allocation of %zu bytes failed\n",
1137 hca
->name
, sizeof *device_attr
);
1138 goto device_init_failed
;
1141 result
= ib_query_device(hca
, device_attr
);
1143 printk(KERN_WARNING
"%s: ib_query_device failed (ret = %d)\n",
1146 goto device_init_failed
;
1148 priv
->hca_caps
= device_attr
->device_cap_flags
;
1152 if (priv
->hca_caps
& IB_DEVICE_UD_IP_CSUM
) {
1153 set_bit(IPOIB_FLAG_CSUM
, &priv
->flags
);
1154 priv
->dev
->features
|= NETIF_F_SG
| NETIF_F_IP_CSUM
;
1158 * Set the full membership bit, so that we join the right
1159 * broadcast group, etc.
1161 priv
->pkey
|= 0x8000;
1163 priv
->dev
->broadcast
[8] = priv
->pkey
>> 8;
1164 priv
->dev
->broadcast
[9] = priv
->pkey
& 0xff;
1166 result
= ib_query_gid(hca
, port
, 0, &priv
->local_gid
);
1168 printk(KERN_WARNING
"%s: ib_query_gid port %d failed (ret = %d)\n",
1169 hca
->name
, port
, result
);
1170 goto device_init_failed
;
1172 memcpy(priv
->dev
->dev_addr
+ 4, priv
->local_gid
.raw
, sizeof (union ib_gid
));
1174 result
= ipoib_dev_init(priv
->dev
, hca
, port
);
1176 printk(KERN_WARNING
"%s: failed to initialize port %d (ret = %d)\n",
1177 hca
->name
, port
, result
);
1178 goto device_init_failed
;
1181 INIT_IB_EVENT_HANDLER(&priv
->event_handler
,
1182 priv
->ca
, ipoib_event
);
1183 result
= ib_register_event_handler(&priv
->event_handler
);
1185 printk(KERN_WARNING
"%s: ib_register_event_handler failed for "
1186 "port %d (ret = %d)\n",
1187 hca
->name
, port
, result
);
1191 if (priv
->dev
->features
& NETIF_F_SG
&& priv
->hca_caps
& IB_DEVICE_UD_TSO
)
1192 priv
->dev
->features
|= NETIF_F_TSO
;
1194 result
= register_netdev(priv
->dev
);
1196 printk(KERN_WARNING
"%s: couldn't register ipoib port %d; error %d\n",
1197 hca
->name
, port
, result
);
1198 goto register_failed
;
1201 ipoib_create_debug_files(priv
->dev
);
1203 if (ipoib_cm_add_mode_attr(priv
->dev
))
1205 if (ipoib_add_pkey_attr(priv
->dev
))
1207 if (ipoib_add_umcast_attr(priv
->dev
))
1209 if (device_create_file(&priv
->dev
->dev
, &dev_attr_create_child
))
1211 if (device_create_file(&priv
->dev
->dev
, &dev_attr_delete_child
))
1217 ipoib_delete_debug_files(priv
->dev
);
1218 unregister_netdev(priv
->dev
);
1221 ib_unregister_event_handler(&priv
->event_handler
);
1222 flush_scheduled_work();
1225 ipoib_dev_cleanup(priv
->dev
);
1228 free_netdev(priv
->dev
);
1231 return ERR_PTR(result
);
1234 static void ipoib_add_one(struct ib_device
*device
)
1236 struct list_head
*dev_list
;
1237 struct net_device
*dev
;
1238 struct ipoib_dev_priv
*priv
;
1241 if (rdma_node_get_transport(device
->node_type
) != RDMA_TRANSPORT_IB
)
1244 dev_list
= kmalloc(sizeof *dev_list
, GFP_KERNEL
);
1248 INIT_LIST_HEAD(dev_list
);
1250 if (device
->node_type
== RDMA_NODE_IB_SWITCH
) {
1255 e
= device
->phys_port_cnt
;
1258 for (p
= s
; p
<= e
; ++p
) {
1259 dev
= ipoib_add_port("ib%d", device
, p
);
1261 priv
= netdev_priv(dev
);
1262 list_add_tail(&priv
->list
, dev_list
);
1266 ib_set_client_data(device
, &ipoib_client
, dev_list
);
1269 static void ipoib_remove_one(struct ib_device
*device
)
1271 struct ipoib_dev_priv
*priv
, *tmp
;
1272 struct list_head
*dev_list
;
1274 if (rdma_node_get_transport(device
->node_type
) != RDMA_TRANSPORT_IB
)
1277 dev_list
= ib_get_client_data(device
, &ipoib_client
);
1279 list_for_each_entry_safe(priv
, tmp
, dev_list
, list
) {
1280 ib_unregister_event_handler(&priv
->event_handler
);
1281 flush_scheduled_work();
1283 unregister_netdev(priv
->dev
);
1284 ipoib_dev_cleanup(priv
->dev
);
1285 free_netdev(priv
->dev
);
1291 static int __init
ipoib_init_module(void)
1295 ipoib_recvq_size
= roundup_pow_of_two(ipoib_recvq_size
);
1296 ipoib_recvq_size
= min(ipoib_recvq_size
, IPOIB_MAX_QUEUE_SIZE
);
1297 ipoib_recvq_size
= max(ipoib_recvq_size
, IPOIB_MIN_QUEUE_SIZE
);
1299 ipoib_sendq_size
= roundup_pow_of_two(ipoib_sendq_size
);
1300 ipoib_sendq_size
= min(ipoib_sendq_size
, IPOIB_MAX_QUEUE_SIZE
);
1301 ipoib_sendq_size
= max(ipoib_sendq_size
, max(2 * MAX_SEND_CQE
,
1302 IPOIB_MIN_QUEUE_SIZE
));
1303 #ifdef CONFIG_INFINIBAND_IPOIB_CM
1304 ipoib_max_conn_qp
= min(ipoib_max_conn_qp
, IPOIB_CM_MAX_CONN_QP
);
1307 ret
= ipoib_register_debugfs();
1312 * We create our own workqueue mainly because we want to be
1313 * able to flush it when devices are being removed. We can't
1314 * use schedule_work()/flush_scheduled_work() because both
1315 * unregister_netdev() and linkwatch_event take the rtnl lock,
1316 * so flush_scheduled_work() can deadlock during device
1319 ipoib_workqueue
= create_singlethread_workqueue("ipoib");
1320 if (!ipoib_workqueue
) {
1325 ib_sa_register_client(&ipoib_sa_client
);
1327 ret
= ib_register_client(&ipoib_client
);
1334 ib_sa_unregister_client(&ipoib_sa_client
);
1335 destroy_workqueue(ipoib_workqueue
);
1338 ipoib_unregister_debugfs();
1343 static void __exit
ipoib_cleanup_module(void)
1345 ib_unregister_client(&ipoib_client
);
1346 ib_sa_unregister_client(&ipoib_sa_client
);
1347 ipoib_unregister_debugfs();
1348 destroy_workqueue(ipoib_workqueue
);
1351 module_init(ipoib_init_module
);
1352 module_exit(ipoib_cleanup_module
);